Research map / Image generation / Autoregressive and latent generation
Effective and Efficient Masked Image Generation Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Unifies masked image generation (MaskGIT/MAR) with masked diffusion models and tunes training and sampling design choices (exp mask schedule, w(t)=1, MAE architecture, CFG with mask, CFG time interval, DPM-Solver) to produce eMIGM, which matches REPA on ImageNet 256 with fewer NFEs and beats EDM2 on ImageNet 512.
Problem
Masked image generation models and masked diffusion models were developed separately, and their design space for training and sampling (mask schedule, loss weighting, architecture, guidance, sampler) is underexplored, leaving a gap in both quality and sampling efficiency relative to continuous diffu
Approach
Shows MaskGIT and MAR losses can be written as special cases of the masked diffusion model (MDM) loss with different masking distributions, weights, and conditional models, then uses this unified view to systematically tune the design space. Key additions are an exponential mask schedule with unit weighting, CFG with a mask token in place of a fake class token, and applying classifier-free guidanc
Key results
- On ImageNet 256x256 with 128x1.4 NFE, eMIGM-H reaches FID 1.57, comparable to REPA's 1.42 while using about 42% of its NFE.
- With similar NFE and parameters, eMIGM-B (16x1.2 NFE, 208M) reaches FID 2.79 vs VAR-d16 at 3.30 (310M).
- On ImageNet 512x512, eMIGM-L at 64x1.25 NFE reaches FID 1.77, slightly beating EDM2-XXL's 1.81 with fewer parameters (478M vs 1.5B).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase model size from eMIGM-B to eMIGM-L | eMIGM · ImageNet 256×256 | FID: 2.79 -> 2.22 | Helped |
| Increase sampling steps from 16×1.2 to 128×1.4 | eMIGM-L · ImageNet 256×256 | FID: 2.22 -> 1.72 | Helped |
| Use the exp mask schedule instead of the linear schedule (ablation) | eMIGM · ImageNet 256×256, epoch 400 | FID: 24.96 -> 11.90 | Helped |
| Use the log-exp mask schedule instead of the exp schedule (ablation) | eMIGM · ImageNet 256×256, epoch 400 | FID: 11.90 -> 9.91 | Helped |
| Use w(t)=1 instead of w(t)=γ′t/γt (ablation) | masked image generation model · ImageNet generation | FID | Helped |
| Use the MAE encoder-decoder architecture instead of a single-encoder transformer (ablation) | masked image generation model · ImageNet generation | FID | Helped |
| Set training time truncation to t_min=0.2 instead of t_min=0 (ablation) | masked image generation model · ImageNet generation | FID | Helped |
| Increase training time truncation to t_min=0.4 instead of t_min=0 (ablation) | masked image generation model · ImageNet generation | FID | Hurt |
| Replace the fake class token with a mask token for unconditional generation (ablation) | CFG for masked image generation · ImageNet generation | FID | Helped |
| Use the exp sampling mask schedule instead of cosine or linear (ablation) | eMIGM · ImageNet generation with 8 or 16 mask prediction steps | FID | Helped |
| Use DPM-Solver instead of DDPM sampling (ablation) | eMIGM diffusion loss · ImageNet generation with 8 mask prediction steps | FID: 10.6 -> 6.6 | Helped |
| Apply classifier-free guidance only within the time interval [0.1, 0.3] (ablation) | eMIGM sampling · ImageNet generation | NFE: 16 × 2; 16 + 16 × (0.3 - 0.1) | Helped |
Surprising findings
- Using a fake class token for unconditional CFG hurts conditional generation performance of MDM, and replacing it with a mask token helps.
- DPM-Solver outperforms DDPM at low mask prediction steps (FID 6.6 vs 10.6 at 8 steps) without temperature tuning.
Limitations
- eMIGM-H on ImageNet 256 is not better than REPA's FID; REPA's result needs self-supervised features but has lower FID at 1.42.
- On ImageNet 512, EDM2-XXL with guidance interval achieves better FID (1.40) with more parameters.
- Log-exp schedule and other mask schedules were only partially explored and left for future work.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet 256x256 ImageNet 512x512
Metrics: FID sFID Inception Score Precision Recall
Baselines: VAR REPA MAR MaskGIT EDM2 DiT-XL/2 SiT-XL StyleGAN-XL Large-DiT DiffiT
Builds on: MaskGIT MAR masked diffusion models (MDM) RADD DPM-Solver MAE
Closest papers by meaning
- Not All Parameters Matter: Masking Diffusion Models for Enhancing Generation Ability (CVPR 2025 · similarity 0.89)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.89)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.88)
- Diffusion Models Without Attention (CVPR 2024 · similarity 0.87)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.87)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.